【时间序列 - 03】ARIMA、ARIMA-ANN(模型融合)

前言:

本文主要介绍 ARIMA 及其模型融合 ARIMA-ANN。由于目前对ARIMA还不是很熟悉,先占个坑位,后续有深入学习,再继续完善,望见谅。


Part1:ARIMA

基本概念

  • ARIMA:Auto Regressive Integrated Moving Average。

  • ARIMA(p,d,q),其中:d 是差分的阶数,用来得到平稳序列;AR是自回归,p(时序数据本身的滞后数)为相应的自回归项;MA为移动平均,q(预测误差的滞后数)为相应的移动平均项数。

  • 模型的输入:历史数据;模型的输出:预测数据;- -

特点

  • 解决:“随机过程的特征随着时间变化而非固定”,“导致时间序列非平稳的原因是随机而非确定”。

  •  

ARIMA 模型的基本流程

  1. 数据可视化,根据历史数据判断时间序列的平稳性。

  2. 对非平稳的时间序列数据,做差分(d),得到平稳序列。

  3. 建立合适的模型: - 平稳化处理后,若偏自相关函数是截尾的,而自相关函数是拖尾的,则建立AR模型; - 若偏自相关函数是拖尾的,而自相关函数是截尾的,则建立MA模型; - 若偏自相关函数和自相关函数均是拖尾的,则序列适合ARMA模型。

  4. 模型的阶数在确定之后,对ARMA模型进行参数估计,比较常用是最小二乘法进行参数估计。

  5. 假设检验,判断残差序列是否为白噪声序列。

  6. 利用已通过检验的模型进行预测。

Reference


Part2:ARIMA-ANN

Panigrahi S, Behera H S, Abraham A. A Fuzzy Filter Based Hybrid ARIMA-ANN Model for Time Series Forecasting[J]. 2013.

摘要:

  • ARIMAlinear model for economic time series forecasting

  • ANNs:capture the complex economic relationships with a variety of patterns

  • Integrating the advantages of ARIMA and ANNs in modeling the linear and nonlinear behaviors;

ARIMA 的优缺点:

  • ARIMA适合处理具有线性相关的时间序列:It assumed that there exists a linear correlation structure among the time series values.

  • 现实生活中的时间序列往往是非线性的关系,而ARIMA处理类似的非线性时间序列的效果往往不佳。

  • ACF:atuocorrelation function;

  • PACF:partial atuocorrelation function;

ANNs 的优缺点:

  • ANNs can function in simple pattern recognition and can be applied to a wide range of application areas

  • ANNs mapping process can cover problems of a greater range of complexity as well as they are superior to other approaches with their powerful, fl exible and easy operation;

  • 1)one of the major advantages of NNs:flexible of modeling nonlinear situations

  • 2)NNs model is constructed adaptively based on the features manifested in the data;

模型融合(ARIMA-ANN):

  • Makridakis et al. (1982) 表示:对于著名的“M-competetion”问题,融合更多的模型能够提升预测性能;

  • Both theoretical and empirical findings suggest that combining different methods is an effective and effi cient way to improve forecasting performance (Palm and Zellner, 1992; Pelikan et al., 1992;

  • Wang and Leu (1996) put forward a hybrid model to forecast the mid-term price trend of the Taiwan stock exchange weighted stock index, which was a recurrent neural network trained by features extracted from ARIMA analyses. (ARIMA提取特征,然后给RNN训练?)

  • 大部分融合模型都使用了ANNs;

ARIMA-ANN:

  • ARIMA、ANNs:non-parameteric techniques;similar in attempting to make appropriate internal representations of time series data;

  • 已有研究证明:融合ARIMA和ANNs能够取得更好地效果;(Kohza di et al., 1996; ElKateb et al. , 1998; Ho et al., 2002)

预测模型的评价指标:

  • MSE:Mean Square Error

  • MAE:Mean Absolute Error

  • MAPE:Mean Absolute Percentage Error

论文中实验的数据集:

  • the Wolf's sunspot data

  • the Canadian lynx data

  • the IBM stock price data

实现代码:

https://github.com/Kanav123/ArimaAnnHybrid

猜你喜欢

转载自blog.csdn.net/Houchaoqun_XMU/article/details/81462865